MultiESC

Alignment of MultiESC

Our implementation is based on MultiESC.

Data Download

Download NRC_VAD.txt, train.txt, valid.txt and test.txt from MultiESC. Put them in the folder data.

Data Preprocessing & Base Model Training

Following the instructions in MultiESC.

Aligned Model Training

Run:

CUDA_VISIBLE_DEVICES=0,1 nohup python align_MultiESC.py --data_type=8 --model_name_or_path=./final_output/lwg_whlookahead_generate --learning_rate=5e-5 --lr2=1e-4 --with_cause --with_strategy --lookahead --model_type=1 --candidate_num=10 --preference_model_dir ../preference_modeling/output/esc_d-PM_23-0527-2227_fold1 --per_device_train_batch_size=6 > align_MultiESC 2>&1 &

Change --preference_model_dir and --model_name_or_path to the path of the preference model checkpoint folder and the base model checkpoint folder, respectively.